講演情報
[AAS08-P20]Integration of Satellite Embeddings and Machine Learning for Enhanced O3 Estimation via Bayesian Maximum Entropy
*Yi Lun Hung1、Hwa Lung Yu1 (1.Department of Bioenvironmental Systems Engineering, National Taiwan University)
キーワード：
Air Quality Prediction、Satellite Remote Sensing、Machine Learning、Bayesian Maximum Entropy、Ozone Concentration、Spatial Estimation
Ground-level ozone (O3) poses substantial risks to human health and ecosystems, motivating the need for accurate spatiotemporal prediction. Bayesian Maximum Entropy (BME) offers a principled framework to fuse hard data (in situ observations) with soft data (probabilistic estimates), yet its performance is strongly influenced by the quality of the soft-data inputs. This study enhances O3 estimation by integrating AlphaEarth foundation-model satellite embeddings with machine learning to generate high-quality soft data for BME-based spatiotemporal mapping.
We propose a three-stage workflow. First, we extracted 64-band AlphaEarth embeddings at 110 ground monitoring locations across Taiwan for 2019–2024, and incorporated digital elevation model (DEM) information as an additional predictor. Due to spatiotemporal matching and quality-control constraints, 88 paired samples were available for model training. Second, to address the small-sample, high-dimensional setting (65 predictors), we applied principal component analysis (PCA), reducing the 64 embedding bands to 7 principal components while retaining 83.8% of the variance. We benchmarked eight regression models using 5-fold cross-validation, and selected an optimized Gradient Boosting Regressor (n_estimators = 100, learning_rate = 0.01, max_depth = 5) as the final soft-data generator. Third, we produced 500 spatially distributed soft-data points over Taiwan, each accompanied by uncertainty summaries (predictive mean ± standard deviation), and integrated them with 110 hard-data locations in BME to estimate O3 on a 50 × 50 × 10 spatiotemporal grid (25,000 points).
Model performance improved substantially through successive optimization. A baseline model using all 64 bands with Random Forest exhibited severe overfitting (RMSE = 2.692 ppb; R2 = −0.182). After PCA, performance improved (RMSE = 2.104 ppb; R2 = 0.278). The final PCA + Gradient Boosting configuration achieved RMSE = 1.873 ppb, R2 = 0.428, relative error = 6.38%, and RMSE% = 14.98%, representing a 30.4% reduction in RMSE relative to the baseline.
Overall, the results indicate that (I) PCA effectively mitigates the curse of dimensionality under limited sample sizes, (II) Gradient Boosting outperforms Random Forest via sequential error correction for this satellite-embedding task, (III) AlphaEarth embeddings capture rich environmental context beyond traditional remote-sensing indices, and (IV) BME successfully integrates machine-learning-derived soft data with uncertainty quantification for spatiotemporal ozone estimation. Future work will incorporate temporal predictors, ensemble uncertainty strategies, and independent monitoring datasets to further strengthen robustness and generalization.
We propose a three-stage workflow. First, we extracted 64-band AlphaEarth embeddings at 110 ground monitoring locations across Taiwan for 2019–2024, and incorporated digital elevation model (DEM) information as an additional predictor. Due to spatiotemporal matching and quality-control constraints, 88 paired samples were available for model training. Second, to address the small-sample, high-dimensional setting (65 predictors), we applied principal component analysis (PCA), reducing the 64 embedding bands to 7 principal components while retaining 83.8% of the variance. We benchmarked eight regression models using 5-fold cross-validation, and selected an optimized Gradient Boosting Regressor (n_estimators = 100, learning_rate = 0.01, max_depth = 5) as the final soft-data generator. Third, we produced 500 spatially distributed soft-data points over Taiwan, each accompanied by uncertainty summaries (predictive mean ± standard deviation), and integrated them with 110 hard-data locations in BME to estimate O3 on a 50 × 50 × 10 spatiotemporal grid (25,000 points).
Model performance improved substantially through successive optimization. A baseline model using all 64 bands with Random Forest exhibited severe overfitting (RMSE = 2.692 ppb; R2 = −0.182). After PCA, performance improved (RMSE = 2.104 ppb; R2 = 0.278). The final PCA + Gradient Boosting configuration achieved RMSE = 1.873 ppb, R2 = 0.428, relative error = 6.38%, and RMSE% = 14.98%, representing a 30.4% reduction in RMSE relative to the baseline.
Overall, the results indicate that (I) PCA effectively mitigates the curse of dimensionality under limited sample sizes, (II) Gradient Boosting outperforms Random Forest via sequential error correction for this satellite-embedding task, (III) AlphaEarth embeddings capture rich environmental context beyond traditional remote-sensing indices, and (IV) BME successfully integrates machine-learning-derived soft data with uncertainty quantification for spatiotemporal ozone estimation. Future work will incorporate temporal predictors, ensemble uncertainty strategies, and independent monitoring datasets to further strengthen robustness and generalization.